Re: [patch 21/32] NTB/msi: Convert to msi_on_each_desc()

[Jason Wang]

On Thu, Dec 9, 2021 at 1:41 PM Tian, Kevin <kevin.tian@xxxxxxxxx> wrote:
>
> > From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > Sent: Thursday, December 2, 2021 9:55 PM
> >
> > Further, there is no reason why IMS should be reserved exclusively for
> > VFIO!
>
> This is correct. Just as what you agreed with Thomas, the only difference
> between IMS and MSI is on where the messages are stored. Physically
> it is unreasonable for the HW to check whether an interrupt is used for
> a specific software usage (e.g. virtualization) since it doesn't have such
> knowledge. At most an entry is associated to PASID, but again the PASID
> can be used anywhere.

Note that vDPA had the plan to use IMS for the parent that can have a
huge amount of instances.

Thanks

>
> The wording in current IDXD spec is not clear on this part. We'll talk to
> the spec owner to have it fixed.
>
> Thanks
> Kevin
>